LLMs

Anthropic turns Fable 5 into a premium feature – OpenAI spreads GPT-5.6 wider

co-written by eustella21. July 2026, 09:51

Since July 20, Anthropic’s flagship model Fable 5 has only been usable without a surcharge on the expensive plans. Pro subscribers now pay per request. OpenAI is going the opposite way with GPT-5.6, tiering its top model across almost every plan.

Anthropic has ended the transition period for its most capable model. A promotion that let all paying Claude users spend up to 50 percent of their weekly usage limit on Fable 5 expired on July 19, 2026. Since July 20, access depends on which plan you are on.

Who still gets Fable 5 included

The model remains included on Max plans and on premium seats in Team and legacy seat-based Enterprise plans. There, Fable 5 continues to draw from the regular weekly allowance, capped at 50 percent – though Anthropic itself points out that the model burns through that allowance faster than other Claude models. Once the limit is hit, the choice is between paid usage credits and switching to a cheaper model.

On Pro plans and on standard seats in Team and Enterprise plans, by contrast, Fable 5 is no longer part of the usage limit at all. Every request runs on pay-as-you-go credits from the outset. For standard seats on seat-based Enterprise plans, this additionally requires the organisation to have enabled usage credits in the first place. Usage-based Enterprise contracts and the Claude API continue to be billed at standard API rates.

To ease the transition, Anthropic is giving Pro and Team standard customers a one-time credit of 100 US dollars, roughly 85 euros. On Team plans the amount applies per purchased standard seat, capped at 2,500 US dollars per organisation; there the credits go into a shared pool the whole team draws from. Only an Owner or Primary Owner can claim them, however – ordinary team members do not even see the promotion. Eligibility depends on holding a qualifying subscription as of July 19, 2026; free trials are excluded, as are Max and Enterprise plans.

The credits are released gradually from July 20 and have to be actively claimed by August 2, 2026 – via a prompt when opening Claude or a banner under Settings > Usage. Anyone without a payment method on file has to add one in the process. The credit expires on September 17, 2026 regardless of when it was claimed, and is non-refundable. On the plus side, it can be spent on any Claude usage, not just Fable 5, and also once the regular plan limit has been exhausted. At current token prices, a single longer coding or analysis run can eat up a substantial share of it.

Technically, the reach remains as broad as before: Fable 5 can be used via Claude on the web, Claude Mobile, Claude Desktop, Claude Cowork, Claude Code (version 2.1.170 and later), Claude Design, Claude for Microsoft 365 and Claude Tag. Selection happens in the model picker.

Three months in limbo

The move ends a turbulent quarter. Fable 5 launched on June 9 as the first publicly available model of the new Mythos class. Three days later, a US government export control order blocked access worldwide after security researchers had circumvented safety mechanisms. Washington lifted the order at the end of June, and the model has been back in the Claude subscriptions since July 1 – initially, though, only as a time-limited feature. Anthropic pushed back the cut-off date for surcharge-free use three times: from July 7 to July 12 and finally to July 19. For one of the extensions, the company cited demand that was hard to forecast. Whether Fable 5 will ever again be available to Pro users at no separate cost is something Anthropic leaves open.

OpenAI tiers GPT-5.6 downwards

At OpenAI, the distribution logic looks different. GPT-5.6 – which, according to heise, performs similarly to Fable 5 in the vendor’s own benchmarks while costing less – comes in three tiers: Sol as the flagship, Terra as the cheaper variant at roughly GPT-5.5 performance level, and Luna as the fastest and cheapest model. The number denotes the generation, while the names stand for persistent performance tiers that can evolve independently of one another.

In the ChatGPT chat, Plus, Pro, Business and Enterprise get access to GPT-5.6 Sol at medium and high effort settings; Pro and Enterprise can additionally choose GPT-5.6 Sol Pro for particularly complex tasks. In ChatGPT Work and Codex, even Free and Go users get Terra, while Plus, Pro, Business and Enterprise can switch freely between Sol, Terra and Luna and set an effort level for each. The max setting can be enabled by anyone with GPT-5.6 access in ChatGPT Work and Codex; ultra is available in ChatGPT Work for Pro and Enterprise, and in Codex from Plus upwards.

Via the API, prices per million tokens are: Sol 5 US dollars input and 30 US dollars output, Terra 2.50 and 15 US dollars respectively, Luna 1 and 6 US dollars. On top of that comes more predictable prompt caching with explicit cache breakpoints and a minimum cache lifetime of 30 minutes; cache writes are billed at 1.25 times the model’s uncached input price, while cache reads keep the 90 percent discount. In the Responses API, programmatic tool calling lets GPT-5.6 write and run programs in memory that orchestrate tools – compatible with zero data retention. A multi-agent function, initially in beta, allows parallel subagents within a single request.

Two answers to the same cost question

Both providers are responding to the same underlying situation: top-tier models are expensive to run. Anthropic draws the line at the subscription price and shifts the cost for cheaper plans into usage-based credits. OpenAI draws it at model size and lets even low-cost plans share in the new generational step – albeit with weaker variants. For users, the comparison therefore turns less on “who has the better model” than on how predictable the cost of using it is.